Deep Physics-Informed Extreme Learning Machines for Orbit Determination

Accurate and robust Orbit Determination (OD) is foundational for numerous space robotics applications, including autonomous docking, planetary rover navigation, and aerial probe control. Traditional approaches such as Weighted Least Squares and the Extended Kalman Filter are widely used due to their statistical reliability, although they can suffer from sensitivity to poor choices of the initial conditions. Recent advances in Machine Learning, particularly Physics-Informed models, offer promising alternatives [1], [2], [3], [4], [5]. Building on earlier work applying Physics-Informed Extreme Learning Machines (PIELMs) to OD [6], [7], a similar framework is introduced here that incorporates orbital dynamics directly into the learning process. This approach is further extended with a Deep PIELM variant featuring autoencoder-based hidden layers [8] to enhance representational capacity and robustness. Comparative results demonstrate the benefits of these deeper architectures in improving OD performance while keeping training times low, positioning them as effective complements to classical techniques.